Predicting Drug Resistance Phenotype of Protein Variants in Colorectal Cancer

Colorectal cancer (CRC) is driven by a complex landscape of somatic mutations that disrupt key signaling pathways involved in cell proliferation, survival, and differentiation. Among these, mutations in BRAF, PTEN, KRAS, MEK1, and PIK3CA play pivotal roles in therapy response and disease progression. Understanding how these mutations alter protein structure and function at the molecular level is essential for improving targeted cancer therapies. In this dissertation, I investigated the structural and dynamic consequences of oncogenic mutations in these five CRC-associated proteins using an integrative framework, the Molecular Dynamic Phenotype Predictive Model(MDPPM), which combines replica exchange with solute tempering (REST2) molecular dynamics simulations and machine learning–based predictive modeling. To build a comprehensive dataset of clinically relevant variants, we systematically curated drug-resistant, drug-sensitive, and variants of uncertain significance (VUS) from publicly available cancer databases, including UniProt, CIViC, ClinVar, the JAX Cancer Knowledgebase, and TCGA. For BRAF, I identified dihedral angle features that distinguish drug-resistant from drug-sensitive variants, revealing conformational mechanisms underlying resistance to dabrafenib and vemurafenib. Similar structural analyses were applied to PTEN, KRAS, and MEK1, uncovering mutation-induced shifts that may disrupt their regulatory roles in the PI3K/AKT and MAPK pathways. Simulations of PIK3CA mutations are currently ongoing to complete the integrated analysis.This work provides new insights into the structural mechanisms that contribute to drug resistance in several key signaling proteins. By combining molecular dynamics simulations with interpretable machine learning, this study demonstrates a practical approach for identifying structural features that distinguish drug-sensitive from resistant variants. A major contribution of this work is the ability to predict the resistance behavior of variants of unknown significance, which remains a difficult problem in the field. While many existing tools rely on evolutionary conservation or general pathogenicity scores, they often fail to capture the local conformational effects caused by specific mutations. In contrast, the methods developed here take into account the dynamic structural context of each protein and offer more accurate predictions. To further evaluate the generalizability of this framework beyond cancer-related proteins, I extended the simulation–machine learning pipeline to a different biological system involving amylin and its analogs. These peptides interact with the amylin receptor and are implicated in neurodegenerative signaling associated with Alzheimer’s disease. By classifying agonists and antagonists based on their structural features, the same computational approach successfully identified key residues predictive of functional activity, suggesting potential utility for rational peptide design. These findings provide a framework that may support more informed and individualized treatment strategies for patients with colorectal cancer and demonstrate that this approach can be adapted to other protein families relevant to human disease.

Author

• Xie, Longsheng

• Alzheimer’s disease
• amylin
• BRAF
• colorectal cancer
• machine learning
• molecular dynamic simulation

Identifier

hdl:1920/15219

OAI identifier oai:identifier

oai:MARS:1920/15219